Rohin Shah comments on [AN #151]: How sparsity in the final layer makes a neural net debuggable